Benchmarking Data-Driven Material Models on Treloar Dataset

Machine learning is transforming constitutive modeling by enabling the direct learning of material behavior from experimental data, challenging traditional paradigms. With a growing array of ML-based approaches for hyperelasticity, a critical need exists to compare their practical performance. This study addresses that need by benchmarking six prominent frameworks: Generalized-Invariant Constitutive Artificial Neural Networks, Physics-Augmented Neural Networks, Adaptive Material Fingerprinting, and Efficient Unsupervised Constitutive Law Identification & Discovery. These methods were rigorously tested against the classic Treloar experimental dataset, a standard in hyperelasticity. The comparison focused on several key aspects: fitting performance, computational cost, sensitivity to hyperparameters, and ease of implementation. While all methods demonstrated remarkable ability to reproduce the benchmark data, the research refrained from identifying a single "winner." Instead, it meticulously outlined the unique strengths and limitations of each approach, providing practical guidance for their application. The source code and data are publicly available.